/SRCNN-RDN

An implementation of SRCNN

Primary LanguagePython

Image Super Resolution Using SRCNN and ResDenseNet

Overview

Single image super-resolution is a challenging image processing task, which aims at recovering a high- resolution image from a single low-resolution image.

In this project, we implement two supervised CNN structures to solve the super-resolution problem, which are the super-resolution convolutional neural network (SRCNN) and the residual dense network (RDN). To improve the model training efficiency, we optimize the SRCNN model with batch normalization.

SRCNN Model

SRCNN

RDN Model

RDN_1 RDN_2

File structure

  • SRCNN folder: all the code, test and training results using SRCNN model.
  • RDN folder: all the code, test and training results using RDN model.

Results

Simulation results show that our SRCNN, SRCNN+BN and RDN models all improve the quality of output images. Among them, SRCNN has the best PSNR and SSIM performance, but RDN has great potential to achieve greater performance by increasing the number of training epoches and the size of training set.